JPH06100918B2 - Voice recognizer - Google Patents

Description

Description: (a) Technical Field of the Invention The present invention recognizes an input feature parameter time series of a speaker's input voice by comparing it with a plurality of registered feature parameter time series registered in advance. The present invention relates to a voice recognition device.

(B) Background of Technology In recent years, as a voice recognition device, input voices of a limited number of registered speakers are converted into appropriate input feature parameter time series and registered in a registered voice dictionary to input voices of unknown speakers. A speaker voice recognition device that determines whether a particular speaker's input voice is identified by comparing it with a registered voice dictionary, or an input voice of a specific speaker's word is converted into an appropriate input feature parameter time series. Convert and register in the registered voice dictionary, collate the input voice uttered by the specific speaker you want to recognize with the registered voice dictionary, and display the recognition result as characters With the progress of voice recognition technology, such as a specific speaker voice recognition device, It has expanded to the field of voice interaction with machines.

As a voice recognition method, unknown input voice is frequency analyzed, the analog signal of the analysis result is converted into a digital signal, the converted digital signal is time-series, the voice section is determined by a threshold value that determines the voice section, and each phoneme is Is extracted as an input feature parameter time series that represents the characteristics of, and the extracted input feature parameter time series is collated with a plurality of registered feature parameter time series that are registered in advance, and the closest distance of the collation result is determined. It is configured to select and obtain a recognition result of an unknown input voice. Since the format of the characteristic parameter time series, the registration method and the matching selection method, etc. affect the difficulty of input speech recognition and the recognition rate, various methods are studied. ing.

(C) Conventional Technology and Problems Conventional speech recognition methods have a registered voice dictionary in which a plurality of input feature parameter time series of the input voice uttered by a specific speaker are registered, and the beginning and ending of the input voice to be recognized are identified. The input feature parameter time series of the determined voice section is compared with the plurality of registered feature parameter time series of the registered voice dictionary, and the recognition result is output. Conventionally, after cutting out one or more voice sections from one input voice, all input feature parameter time series of each voice section are used, and the distance is calculated for each of a plurality of registered feature parameter time series registered in advance. I was doing calculations.

The configuration of this type of voice recognition system will be described. FIG. 1 shows a circuit block diagram of a conventional speech recognition apparatus. The procedure for registering the voice of the speaker in advance is to input one input voice uttered by the speaker from the microphone 1 and input the input voice in the band filter group 2 in the voice band 200 Hz to 5 KHz.
Divide the degree into 10 to 20 channel filter groups, and
Each channel filter output is taken out at a period of 30 ms, converted by the characteristic parameter time series extraction unit 3 into the input characteristic parameter time series of digital information, and stored in the input characteristic parameter time series buffer 4. The input feature parameter time-series stored in the input feature parameter time-series buffer 4 is a registration feature representing the feature of the phoneme of the input voice by extracting a voice segment with a threshold value that determines the voice segment at the beginning and end of the speech by the voice segment extraction circuit 5. It is registered in the registered characteristic parameter time series dictionary unit 6 as a parameter time series. The above procedure is repeated for the number of input voices to be registered, and a plurality of registered characteristic parameter time series are registered in the registered characteristic parameter time series dictionary unit 6.

Next, in the procedure for recognizing the voice of the speaker, the input voice uttered by the speaker is input from the microphone 1, and the input voice is stored in the input feature parameter time series buffer 4 according to the same procedure as described above. The input feature parameter time series stored in the input feature parameter time series buffer 4 is divided by the voice section cutout circuit 5 into a plurality of voice sections according to a threshold value for determining the voice section and cut out. The voice section cutout circuit 5 may be of a type in which only the threshold value operated by the same algorithm is changed or a combination type of circuits in which the algorithms themselves are different.

FIG. 4 shows an example in which a plurality of voice sections are determined by the threshold for determining the voice section in the voice section cutout circuit 5.
Let TL, TM, TH be thresholds of three levels, and one voice, for example,
When "Aomori" is input, "Aomori" at the threshold TL level, "Aomori" at the threshold TM level, and the threshold TH depending on the threshold.
Three voice segments of "omo" are decided by the level. Therefore, when n threshold values are set in FIG. 1, n input feature parameter time series of the voice section are output, and a plurality of registered feature parameter time series dictionary units 6 are registered. The feature parameter time series m and the input feature parameter time series n for each voice section are sequentially matched by the matching selection circuit 7 n × m times, and the matching distance is calculated n × m times to obtain n × m. The closest matching distance is selected from the matching distances of the times and is output to the recognition terminal 8 as a recognition result.

The above is the processing procedure of the conventional voice recognition. In this method, the matching distance calculation amount of the n input feature parameter time series of the voice section and the m registered feature parameter time series is n ×
Since the number of recognitions is m, the calculation amount increases as the number of recognition categories increases, the time required for the matching process increases, and the false recognition rate is large, which is a drawback.

(D) Object of the Invention The present invention aims to solve the above-mentioned conventional drawbacks.

(E) Structure of the invention The above-mentioned object is, according to the present invention, a characteristic parameter time series for converting each channel of the input sound divided into i channels by a band filter into a digitized i number of input characteristic parameter time series. Extraction unit, voice section cutout circuit for cutting out input feature parameter time series of each of i channels into n voice sections using n threshold values, registered feature parameter time series dictionary section for registering output from the circuit In the configuration including, the first selection circuit,
A second selection circuit is provided, and the unknown input speech divided into i channels is converted into the i input characteristic parameter time series digitized by the characteristic parameter time series extraction unit, and each of them is converted by the speech segmentation circuit. n thresholds
I channels cut out into audio segments,
For such i channels, the first selection circuit selects j (j <i) channels each having n speech sections as a coarse input feature parameter time series, and stores them in the registered feature parameter time series dictionary unit. Matched with one of the m registered registration parameter time series in advance,
The second selection circuit has a function of selecting one of n speech sections having the closest matching distance for each matching and outputting m selection results, and the second selection circuit includes a first speech section consisting of a time series of rough input feature parameters. Converted from m outputs from the selection circuit,
Registered features of the registered feature parameter time series dictionary unit corresponding to m or r (r <m) outputs consisting of speech sections including all input feature parameter time series and m outputs in the first selection circuit This is achieved by a voice recognition device having a function of matching and selecting a parameter time series in correspondence with m times or r times.

By providing a plurality of voice section cutout circuits and a two-stage selection circuit according to the present invention, the calculation amount of the matching distance can be reduced as compared with the conventional method, and at the same time, the input feature parameter time series and the registered feature parameter time Series 2
It is similar to collating twice, and recognizes it according to various changes such as the strength and time of the input voice, time length, and accent. Especially, even if the voice is faint due to the physical condition of the speaker on the day, it is recognized with a high recognition rate. There is an advantage that can be done.

(F) Example of Invention An example of the present invention will be described below. FIG. 2 is a circuit configuration block diagram of the speech recognition apparatus according to the present invention. Throughout all the drawings, the same objects are designated by the same reference numerals as those in FIG. 9 is a first selection circuit, 10 is a registered feature parameter time series dictionary unit,
Reference numeral 11 is an AND gate circuit, and 12 is a second selection circuit.

In this circuit configuration, the voice registration processing procedure is the same as the conventional method shown in FIG. The voice recognition processing procedure according to the present invention is the same as that shown in FIG. 1 up to the point where the input voice of the speaker to be recognized is input from the microphone 1 and the input feature parameter time series is stored in the input feature parameter time series buffer 4. This input feature parameter time series is cut out as an input feature parameter time series of n voice sections by setting n threshold values in the voice section cutout circuit 5, and in the first selection circuit 9, channel i of the band filter group 2 is selected. M channels that are registered in advance in the registered feature parameter time series dictionary unit 10 after being converted into rough input feature parameter time series for j channels (where j <i) each having n voice sections. One of the registered feature parameter time series of is sent to the first selection circuit 9 to be compared with the n coarse input feature parameter time series, and one voice section having the closest matching distance is selected from n. Only one voice segment selected from the n voice segments is input to the a side of the AND gate circuit 11. At the same time, the input feature parameter time series having all i channels of the voice section cut out with the same threshold value is input to the b side of the AND gate circuit 11 of the same system. The input feature parameter time series is output from the c side of the AND gate circuit 11 only for the systems input to both sides a and b of the AND gate circuit 11. Then, the other one of the registered characteristic parameter time series of m is compared with the rough input characteristic parameter time series of the n speech sections, the speech section is selected in the same manner as described above, and the AND gate circuit 11 The input feature parameter time series of the selected voice section is output to the c side. Thus, the time series of the coarse input feature parameters of n voice sections are collated with the time series of m registered feature parameters m times, and a total of m voice sections are selected and sequentially selected on the c side of the AND gate circuit 11. Only the input feature parameter time series of the i channel of the selected voice section is output and input to the second selection circuit 12. The input feature parameter time series of the i channels of the m selected voice sections and the m registered feature parameter time series registered in the registered feature parameter time series dictionary unit 10 are sequentially selected by the second selection circuit 12 Is calculated m times, the closest matching distance is selected, and the registered feature parameter time series selected as the recognition result is output to the recognition terminal 8. In the method of the present invention, the calculation amount of the matching distance calculated by the first selection circuit and the second selection circuit is i and j such that (j / i) <(1-1 / n) and j <i. By setting, (nj / i + 1) m <(n × m), and the amount of calculation decreases.

FIG. 3 shows another embodiment of the present invention, in which 13 is a first selection circuit, 14 is a selection input characteristic parameter time series buffer, and 15 is a second selection circuit. The first selection circuit 13 collates the rough input feature parameter time series n for channel j among channels i described in FIG. 2 with the registered feature parameter time series m, and determines the closest speech segment r among the collation distances. Individual pieces (however, r ≦ m) are selected and stored in the selected input feature parameter time series buffer 14. The second selection circuit 15 calculates the matching distance between the respective r voice sections of all the i channels selected and stored, selects the closest matching distance, and outputs it as the recognition result to the recognition terminal 8. The calculation amount of the matching distance calculated by the first selection circuit and the second selection circuit should be (j / i) <(1-r / nm), r ≦ m, j <i
By selecting i, j, r, (j / i) × nm + r <(n × m) is obtained.

As described above, the frequency analysis method using the band filter has been described as the embodiment of the present invention, but the present invention can also be applied to a voice recognition device adopting the LPC analysis method or the like.

(G) Effect of the Invention As described above, by providing a plurality of voice section cutout circuits according to the present invention and a two-stage selection circuit,
In the speech recognition for recognizing the input voice of the speaker, the recognition rate can be improved and the amount of calculation of the matching distance can be reduced because the input voice can be recognized in response to various variations such as the strength, time length, and accent. is there.

[Brief description of drawings]

FIG. 1 is a circuit configuration block diagram of a conventional voice recognition device,
FIG. 2 and FIG. 3 are circuit block diagrams of the voice recognition device according to the present invention, and FIG. 4 is a relational diagram for determining the voice section of the input voice. In the drawing, 1 is a microphone, 2 is a band filter group, 3
Is a feature parameter time series extraction unit, 4 is an input feature parameter time series buffer, 5 is a voice segment cutout circuit, 7 is a matching selection circuit, 8 is a recognition terminal, 9 is a first selection circuit, and 10 is a registered feature parameter time series dictionary. Part, 12 is the second selection circuit, 13
Is a first selection circuit, 14 is a selection input feature parameter time series buffer, and 15 is a second selection circuit.

Claims (1)

[Claims] 1. A feature parameter time series extraction unit for converting each channel of input speech divided into i channels by a band filter into a digitized i number of input feature parameter time series. In a configuration including a voice section cutout circuit that cuts out each input feature parameter time series into n voice sections using n thresholds, and a registered feature parameter time series dictionary unit that registers an output from the circuit, the first selection A circuit and a second selection circuit are provided, and the unknown input speech divided into i channels is converted into i input characteristic parameter time series digitized by the characteristic parameter time series extraction unit, and each is segmented into a voice section. In the circuit, there are i thresholds cut out into n voice segments and i channels are cut out. , The first selection circuit selects j (j <i) channels each having n voice sections as a rough input feature parameter time series, and registers them in the registration feature parameter time series dictionary unit in advance. It has a function of matching one of m feature parameter time series to one destination, selecting one of n voice segments having the closest matching distance for each match, and outputting m selection results. The circuit transforms m outputs from the first selection circuit for a speech segment having a rough input characteristic parameter time series into m or r speech segments including all the input characteristic parameter time series (r < m) and the registered feature parameter time series of the registered feature parameter time series dictionary unit corresponding to the m outputs in the first selection circuit have a function of matching and selecting m times or r times. Characterized by Voice recognition device.